Reproduce “CMT: A Cascade MAR with Topology Predictor for Multimodal Conditional CAD Generation” on OpenResearch

About this paper

Jianyu Wu, Yizhou Wang, Xiangyu Yue, Xinzhu Ma, Jingyang Guo, Dongzhan Zhou, Wanli Ouyang, Shixiang Tang

Researchers introduce CMT (Cascade MAR with Topology Predictor), a framework that enables multimodal conditional generation of Boundary Representation (B-Rep) CAD models by explicitly learning topology through a cascade autoregressive network. The method achieves higher geometric and topological accuracy, along with improved validity, compared to prior art in both unconditional and conditional settings, facilitated by the creation of the large-scale multimodal mmABC dataset.

Key insights

• Generating B-Rep components in a cascade (edges first, then surfaces informed by edges) effectively captures and utilizes the inherent topological priors, leading to more valid models.
• Continuous tokenization coupled with autoregressive generation is crucial for maintaining precision and managing the complexity of B-Rep models.
• A unified encoder can effectively leverage powerful pre-trained multimodal models (like CLIP) to enable robust conditional generation from diverse input modalities.
• Explicitly predicting the topological adjacency between generated edges and surfaces is more efficient and accurate than traditional post-processing for B-Rep reconstruction.

Summary by alphaXiv.